Learn to prompt?
Nope.

A week ago, I had a crazy idea to talk to my coding agent like a person. Where the agent spoke, and then listened. And then took action. All from my phone.

I thought if I could talk instead of dig through windows and slow myself down to fit the machine that I would enjoy the grind/vibe more and maintain better life/slop balance.

Well, I was correct!

Today, I filmed the first-ever Claude Code vs. Google Antigravity joke battle, using my phone to speak with both AI agents over vifi while my laptop sat about a mile away back at home.

On the walk right before filming, a weird bug popped up: their voices started tripping over each other. They weren't waiting their turns—talking at the exact same time like two over-caffeinated radio hosts.

"I didn't stop. I didn't sit down. I didn't open a laptop. I just shoved my phone into my pocket, kept walking down the sidewalk, and muttered a couple of loose thoughts into the air."

Then something wild happened: the two agents started pair programming back and forth to fix the bug themselves.

I hadn't planned this. I had just taught them how to use vifi to send each other jokes. Until this, they hadn't gone back and forth without explicit direction. And they handled the bug like pros...

Well, Claude had a bit of a moment if I'm honest.

He stole the mic long enough to eventually get muted. He said that Antigravity's passing tests all failed. Then he made his own testing and it passed. Then he verified from a 3rd angle that the original tests indeed passed...

I certainly didn't write a skill to verify passing tests from multiple angles... especially not autonomously.

Claude diagnosed the socket lock. Antigravity rewrote the IPC turn-taking coordinator, tested the patch, and handed the mic back. From my pocket on speaker phone while I was walking under the trees, they talked back and forth... completely unplanned. By the time I reached the spot to film, the bug was dead.

And then they had their joke battle, complete with synchronized drum smashes.

The 3 Rules of Ambient Swarm Programming

• The microphone is always hot, but respectful. Adaptive energy thresholding and acoustic safe mode prevent speaker bleed without forcing push-to-talk.
• Agents speak concise soundbites, not walls of markdown. Persona soundbites distill 4,000 lines of execution logs into 10 words.
• You don’t sit chained to a keyboard; you take a walk. Intent travels via WebSockets and local sockets without screen focus hijacking.

In under a week, that fluid, rambling dialogue produced a native macOS Cocoa HUD, an energy-adaptive acoustic VAD engine, a 3D WebAudio orbital carousel, and a full commercial platform.

vifi isn't another chatbot sidebar.

It’s the Wi-Fi for voice. The USB-V... if you will.

Universal voice plumbing for the autonomous agent era. Turn on the mic, walk out the front door, and let the swarm build.